def test_predict_from_model_array(self):
params = nn_params.copy()
params.update({'nclasses': n_classes})
optimizer = GenericSolver(**solver_param)
datasets.dump_svmlight_file(xtr, ytr, tr_f)
datasets.dump_svmlight_file(xte, yte, te_f)
# Train model
clf = MLP(**params)
clf.fit(xtr_arr, validation_data=[xte_arr], solver=optimizer)
y_pred_tr = clf.predict(xtr_arr)
y_pred_te = clf.predict(xte_arr)
# Load from tained model
params = nn_params.copy()
params.update({
'finetuning': True,
'template': None,
'nclasses': n_classes
})
clf = MLP(sname=clf.sname,
repository=clf.model['repository'],
**params)
assert np.array_equal(y_pred_tr, clf.predict(xtr_arr))
assert np.array_equal(y_pred_te, clf.predict(xte_arr))
os_utils._remove_files([tr_f, te_f])
def test_lmdb_creation(self):
params = nn_params.copy()
params.update({'nclasses': n_classes})
# Create dataset
X, Y = datasets.load_digits(return_X_y=True)
X = preprocessing.StandardScaler().fit_transform(X)
x_train, x_test, y_train, y_test = model_selection.train_test_split(
X, Y, test_size=test_size, random_state=seed)
# Save data in .svm format
tr_svm_f, tr_lmdb_f = os.path.abspath('x_train.svm'), os.path.abspath(
'x_train.lmdb')
te_svm_f, te_lmdb_f = os.path.abspath('x_test.svm'), os.path.abspath(
'x_test.lmdb')
vocab_path = os.path.abspath('vocab.dat')
datasets.dump_svmlight_file(x_train, y_train, tr_svm_f)
datasets.dump_svmlight_file(x_test, y_test, te_svm_f)
lmdb_utils.create_lmdb_from_svm(svm_path=tr_svm_f,
lmdb_path=tr_lmdb_f,
vocab_path=vocab_path,
**params)
lmdb_utils.create_lmdb_from_svm(svm_path=te_svm_f,
lmdb_path=te_lmdb_f,
**params)
tr_lmdb = SVMConnector(path=tr_svm_f,
lmdb_path=tr_lmdb_f,
vocab_path=vocab_path)
te_lmdb = SVMConnector(path=te_svm_f, lmdb_path=te_lmdb_f)
optimizer = GenericSolver(solver_type='SGD',
base_lr=0.01,
iterations=100)
clf = MLP(**params)
clf.fit(tr_lmdb, validation_data=[te_lmdb], solver=optimizer)
ytr_prob = clf.predict_proba(tr_lmdb)
acc = metrics.accuracy_score(y_train, ytr_prob.argmax(-1))
assert acc > 0.7
os_utils._remove_files([tr_svm_f, te_svm_f, vocab_path])
os_utils._remove_dirs([tr_lmdb_f, te_lmdb_f])
def test_classification(self):
params = nn_params.copy()
params.update({'nclasses': n_classes})
optimizer = GenericSolver(**solver_param)
datasets.dump_svmlight_file(xtr, ytr, tr_f)
datasets.dump_svmlight_file(xte, yte, te_f)
clfs = [
# array connector without validation set
[xtr_arr, [], MLP(**params)],
[xtr_arr, [], LR(**params)],
# sparse array connector without validation set
[xtr_sparse, [], MLP(**params)],
[xtr_sparse, [], LR(**params)],
# svm connector without validation set
[xtr_svm, [], MLP(**params)],
[xtr_svm, [], LR(**params)],
# array connector with validation set
[xtr_arr, [xte_arr], MLP(**params)],
[xtr_arr, [xte_arr], LR(**params)],
# svm connector with validation set
[xtr_svm, [xte_svm], MLP(**params)],
[xtr_svm, [xte_svm], LR(**params)],
]
for tr_data, te_data, clf in clfs:
clf.fit(tr_data, te_data, optimizer)
y_pred = clf.predict(tr_data)
acc = metrics.accuracy_score(ytr, y_pred)
print(acc)
assert acc > 0.7
os_utils._remove_files([tr_f, te_f])
def create_lmdb_from_svm(svm_path,
lmdb_path,
vocab_path=None,
host='localhost',
port=8085,
nclasses=2,
gpu=True,
tmp_folder=None):
if os.path.exists(lmdb_path):
print("warning: {} exist, overwriting it".format(lmdb_path))
tmp_folder = tempfile.mkdtemp(
prefix="pydd_", dir=tmp_folder) if tmp_folder else tempfile.mkdtemp(
prefix="pydd_")
train_data = SVMConnector(path=svm_path)
optimizer = GenericSolver(solver_type='SGD', base_lr=0.01, iterations=1)
clf = MLP(host=host,
port=port,
nclasses=nclasses,
gpu=gpu,
repository=tmp_folder)
clf.fit(train_data, solver=optimizer)
shutil.move(os.path.join(tmp_folder, "train.lmdb"), lmdb_path)
if vocab_path:
shutil.move(os.path.join(tmp_folder, "vocab.dat"), vocab_path)
# delete service
clf.delete_service(clf.sname, clear='lib')
# delete tmp_folder
shutil.rmtree(tmp_folder)
return lmdb_path, vocab_path
# Parameters
seed = 1337
n_classes = 10
repository = "/tmp/pydd_test"
params = {
"repository": repository,
"port": 8080,
"nclasses": n_classes,
"gpu": True
}
split_params = {"test_size": 0.2, "random_state": seed}
np.random.seed(seed) # for reproducibility
solver = GenericSolver(iterations=1000,
solver_type="SGD",
base_lr=0.01,
gamma=0.1,
stepsize=30,
momentum=0.9,
snapshot=200)
class_weights = [1., 1., 1., 1., 1., 1., 1., 1., 1., 1]
# remove repository if existe else creates it
if os.path.exists(repository):
os_utils._remove_dirs([repository])
os.makedirs(repository)
# create dataset
X, y = datasets.load_digits(n_class=n_classes, return_X_y=True)
X = preprocessing.StandardScaler().fit_transform(X)
xtr, xte, ytr, yte = model_selection.train_test_split(X, y, **split_params)
X, y = datasets.load_digits(n_class=n_classes, return_X_y=True)
X = preprocessing.StandardScaler().fit_transform(X)
x_train, x_test, y_train, y_test = model_selection.train_test_split(X, y, test_size=test_size, random_state=seed)
tr_f = os.path.abspath('x_train.svm')
te_f = os.path.abspath('x_test.svm')
datasets.dump_svmlight_file(x_train, y_train, tr_f)
datasets.dump_svmlight_file(x_test, y_test, te_f)
# train_data = ArrayConnector(x_train, y_train)
# val_data = ArrayConnector(x_train, y_train)
train_data = SVMConnector(tr_f)
val_data = SVMConnector(te_f)
clf = MLP(host=host, port=port, nclasses=n_classes, layers=[100], gpu=gpu)
solver = GenericSolver(iterations=iteration, test_interval=30, solver_type="SGD", base_lr=lr)
clf.fit(train_data, validation_data=[val_data], solver=solver)
clf.predict_proba(train_data)
clf.fit(train_data, validation_data=[val_data], solver=solver)
y_pred = clf.predict_proba(train_data)
clf = LR(host=host, port=port, nclasses=n_classes, gpu=gpu)
solver = GenericSolver(iterations=iteration, solver_type="SGD", base_lr=lr)
clf.fit(train_data, solver=solver)
y_pred = clf.predict_proba(train_data)
clf = XGB(host=host, port=port, nclasses=n_classes)
# logs = clf.fit(train_data, validation_data=[val_data])
os_utils._remove_files([tr_f, te_f])
test_size=test_size,
random_state=seed)
# create and save train.svm and test.svm
tr_f = os.path.abspath('x_train.svm')
te_f = os.path.abspath('x_test.svm')
datasets.dump_svmlight_file(xtr, ytr, tr_f)
datasets.dump_svmlight_file(xte, yte, te_f)
# create connectors
xtr_svm, xte_svm = SVMConnector(tr_f), SVMConnector(te_f)
# train model
params = {'host': host, 'port': port, 'nclasses': nclasses, 'layers': [100]}
optimizer = GenericSolver(solver_type='SGD',
iterations=500,
base_lr=0.1,
snapshot=100)
clf = MLP(sname=sname, repository=model_repo, **params)
clf.fit(xtr_svm, validation_data=[xte_svm, xtr_svm], solver=optimizer)
del clf
# load pre trained model
params = {
'host': host,
'port': port,
'nclasses': nclasses,
'finetuning': True,
'template': None
}
clf = MLP(sname=sname, repository=model_repo, **params)
ytr_pred, yte_pred = clf.predict(xtr_svm), clf.predict(xte_svm)
x_train, x_test, y_train, y_test = model_selection.train_test_split(
X, Y, test_size=test_size, random_state=seed)
# Save data in .svm format
tr_svm_f, tr_lmdb_f = os.path.abspath('x_train.svm'), os.path.abspath(
'x_train.lmdb')
te_svm_f, te_lmdb_f = os.path.abspath('x_test.svm'), os.path.abspath(
'x_test.lmdb')
vocab_path = os.path.abspath('vocab.dat')
datasets.dump_svmlight_file(x_train, y_train, tr_svm_f)
datasets.dump_svmlight_file(x_test, y_test, te_svm_f)
# create lmdb and vocab file
create_lmdb_from_svm(svm_path=tr_svm_f,
lmdb_path=tr_lmdb_f,
vocab_path=vocab_path,
**params)
create_lmdb_from_svm(svm_path=te_svm_f, lmdb_path=te_lmdb_f, **params)
tr_data = SVMConnector(path=tr_svm_f,
lmdb_path=tr_lmdb_f,
vocab_path=vocab_path)
te_data = SVMConnector(path=tr_svm_f, lmdb_path=tr_lmdb_f)
optimizer = GenericSolver(solver_type='SGD', base_lr=0.01, iterations=100)
clf = MLP(**params)
clf.fit(tr_data, validation_data=[te_data], solver=optimizer)
y_pred_lmdb = clf.predict_proba(te_data)
X, y = datasets.load_digits(n_class=n_classes, return_X_y=True)
X = preprocessing.StandardScaler().fit_transform(X)
xtr, xte, ytr, yte = model_selection.train_test_split(X, y, **split_params)
# create and save train.svm and test.svm
tr_f = os.path.abspath('x_train.svm')
te_f = os.path.abspath('x_test.svm')
datasets.dump_svmlight_file(xtr, ytr, tr_f)
datasets.dump_svmlight_file(xte, yte, te_f)
# Define models and class weights
clf = MLP(**params)
solver = GenericSolver(iterations=500,
solver_type="SGD",
base_lr=0.01,
gamma=0.1,
stepsize=30,
momentum=0.9)
# one class weight value for each class
class_weights = [1., 1., 1., 1., 1., 1., 1., 1., 1., 1]
train_data, test_data = SVMConnector(path=tr_f), SVMConnector(path=te_f)
logs = clf.fit(train_data,
validation_data=[test_data],
solver=solver,
class_weights=class_weights,
batch_size=128)
yte_pred = clf.predict(test_data)
report = metrics.classification_report(yte, yte_pred)
print(report)